Heterogeneous Transfer-Learning-Enabled Diverse Metasurface Design

摘要

With the rapid growth in intelligent metasurfaces in the recent years, deeplearning has attracted attention to transform the ways in which metasurfacesare simulated and designed. The unique advantages of deep learning liein the powerful data-driven modality, which allows a computational modelto learn useful information using hierarchically structured layers. Amongthe various successful examples, there are forward and inverse designs.However, such designs are inherently data-hungry. Thus, the data utilizationefficiency must be maximized, and green metasurface design must beachieved. Here, the authors propose heterogeneous transfer learning to allowtransferrable and data-efficient metasurface design. The key to this methodis a flexible network framework, which integrates feature augmentation anddimensionality reduction. The concept is demonstrated through three scenarios,i.e., metasurfaces with different parameterizations, different physicalsizes, and completely different geometries, where the relative error reductionreaches up to 50. Furthermore, an inverse metasurface design is proposed,which combines the forward predicted network and heuristic algorithm. Thiswork considerably reduces the workload on data collection and overcomesthe limitation that previous works only focused on fixed physical structures.The authors have also envisioned a “global metasurface gene bank,” in whichresearchers can freely “withdraw and save data” for various applications.